Leveraging Feedback for Dynamic Execution Optimization Yuxuan Zhang

Format:

Book

Thesis/Dissertation

Author/Creator:

Zhang, Yuxuan, author.

Contributor:

University of Pennsylvania. Computer and Information Science., degree granting institution.

Language:

English

Subjects (All):

0464.

0723.

0984.

Local Subjects:

0464.

0723.

0984.

Physical Description:

1 electronic resource (160 pages)

Contained In:

Dissertations Abstracts International 87-07A

Place of Publication:

Ann Arbor : ProQuest Dissertations and Theses, 2025

Language Note:

English

Summary:

Modern computer systems, characterized by deep and complex software stacks, suffer from a fundamental abstraction gap between applications, system software, and hardware. This gap obscures low-level execution behavior and introduces significant microarchitectural performance bottlenecks-such as memory system inefficiencies. At the same time, intermediary layers such as virtualization, container frameworks, and microservice runtimes introduce additional system-level overheads, including scheduling delays, memory allocation inefficiencies. These challenges are difficult to mitigate through application-level optimizations alone.Although a wide range of hardware and software solutions have been proposed, they often face practical deployment barriers or require disruptive architectural changes. Compiler-based Profile-Guided Optimization (PGO) offers a promising cross-layer approach but relies on static, offline profiling that fails to capture the dynamic and reactive nature of modern workloads. In resource-variable cloud environments, such static profiles quickly become obsolete, resulting in mismatched optimizations and limited performance gains.This dissertation demonstrates that dynamic, cross-layer Feedback-Driven Optimization (FDO) can effectively bridge this performance gap. By leveraging real-time microarchitectural metrics and system-level telemetry, this approach enables applications to adapt continuously to changing hardware conditions, workload characteristics, and resource availability without recompilation or redeployment.To validate this thesis, the dissertation presents three systems that apply FDO across different layers of the computing stack. OCOLOS introduces the first online code layout optimization system for unmodified C/C++ applications, optimizing instruction cache locality and branch prediction behavior directly within running processes to achieve performance gains of up to 2.2×. RPG2 is a software-only system for dynamic data prefetch injection and tuning that mitigates backend memory latency through continuous runtime adaptation, yielding speedups of up to 2.19× while avoiding prefetch-induced slowdowns. Quilt extends FDO principles to serverless computing, using dynamic profiling of system telemetry to guide the intelligent co-location and merging of functions, reducing invocation latency and improving workflow throughput by up to 12.87×.Collectively, these systems demonstrate that leveraging cross-layer runtime feedback enables the construction of self-optimizing applications that evolve alongside the dynamic environments in which they operate. This dissertation contributes both the methodology and empirical evidence showing that cross-layer feedback-driven optimization can transform system observability into actionable adaptation, narrowing the long-standing divide between applications and system/hardware realities

Notes:

Advisors: Angel, Sebastian G. Committee members: Liu, Vincent F.; Devietti, Joseph L.; Lee, Benjamin C.; Litz, Heiner; Khan, Tanvir Ahmed

Source: Dissertations Abstracts International, Volume: 87-07, Section: A.

Ph.D. University of Pennsylvania 2025

Vendor supplied data

Local Notes:

School code: 0175

ISBN:

9798276005157

Access Restriction:

Restricted for use by site license